Fine-tuning vs RAG: choosing the right knob

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-creators-fine-tuning-vs-rag

1. A company notices their chatbot sometimes refuses to answer harmless questions that it should answer. What type of problem is this?

   1. A context problem requiring more data
   2. A knowledge problem requiring fresh information
   3. A hardware problem requiring better GPUs
   4. A behavior problem requiring style adjustment

2. A legal tech startup needs their LLM to cite recent court decisions that didn't exist when the model was trained. What approach directly solves this?

   1. Fine-tuning on legal briefs
   2. RAG with current case law
   3. Temperature adjustment
   4. Reinforcement learning from human feedback

3. What is the primary danger of fine-tuning a model on noisy or messy training data?

   1. The model loses all its original capabilities
   2. The model becomes too slow to respond
   3. The training costs become prohibitive
   4. The model learns incorrect patterns and presents them confidently

4. A team wants their model to respond in a specific JSON format for every API call. Which approach is most appropriate?

   1. Increasing the model temperature
   2. Adding more system prompts at runtime
   3. RAG with JSON schema documents
   4. Fine-tuning with JSON-formatted examples

5. Why do most teams need RAG before they need fine-tuning?

   1. Most problems involve knowledge gaps rather than behavioral ones
   2. RAG is cheaper than fine-tuning in all cases
   3. RAG is easier to implement than fine-tuning
   4. Fine-tuning requires more expensive GPUs

6. A model consistently provides outdated information about a company's current product pricing. What is the root cause?

   1. Behavioral misalignment requiring fine-tuning
   2. Knowledge gap requiring fresh context
   3. System prompt too long
   4. Temperature set too high

7. What does 'evaluation discipline' mean in the context of LLM development?

   1. Letting users decide if the model is good enough
   2. Evaluating models once during initial training
   3. Systematically testing model outputs against defined metrics before and after changes
   4. Evaluating models only after deployment to production

8. A team wants their customer service bot to sound more empathetic while also knowing the latest return policy. What is the minimum approach needed?

   1. Fine-tuning alone
   2. Both RAG and fine-tuning
   3. Neither—system prompts can handle both
   4. RAG alone

9. What is 'time-to-value' primarily measuring in the context of fine-tuning versus RAG?

   1. The latency of retrieval in RAG systems
   2. How long the model takes to generate each response
   3. The computational time required for training
   4. How quickly an approach delivers meaningful improvements

10. A model generates correct-sounding but factually wrong information about a specialized topic. What best describes this failure mode?

    1. Knowledge hallucination requiring RAG
    2. Behavior problem requiring refusal training
    3. Context window overflow
    4. Temperature problem requiring lowering

11. Why can't AI eliminate the need for a real evaluation suite?

    1. Evaluation requires human judgment about whether outputs are actually correct and useful
    2. Evaluation suites are illegal without human oversight
    3. AI models are not advanced enough yet
    4. GPU costs are too high for AI evaluation

12. A startup has clean, well-labeled training data and wants the model to adopt a specific brand voice. Why is this still insufficient without evaluation?

    1. Evaluation measures whether the fine-tuned model actually achieves the desired behavior
    2. Fine-tuning always works regardless of data quality
    3. Clean data is not needed for brand voice tasks
    4. Clean data causes overfitting in most cases

13. What does RAG provide that fine-tuning fundamentally cannot?

    1. Lower inference costs
    2. Fresh information retrieved at query time
    3. Better reasoning capabilities
    4. Improved token limit handling

14. What is the main cost driver that makes fine-tuning more expensive than RAG for most use cases?

    1. Training compute and iteration cycles
    2. Retrieval system infrastructure
    3. API call volume
    4. Prompt engineering effort

15. A team implements RAG but users still complain about incorrect answers. What might be missing in their implementation?

    1. Their model is too small
    2. They used fine-tuning instead
    3. They lack evaluation to verify retrieval quality and answer correctness
    4. They didn't use enough system prompts